System Design Frontier (SDF): What made you start RISC-I project in 1980s? Are there any inside stories you wanna share with us?

David Patterson (DP): I had been working on a research project as a brand new assistant professor, just got my PhD. Probably did not make any sense. I had been on sabisatical at DEC, which was making a more complicated min-computer. As a result of that experience, at Berkeley we were trying to design single-chip computer (microprocessor). When I came back from DEC, realizing it would be very difficult to build this VAX microprocessor, compatible with VAX instruction set on a single-chip. It did not make sense to build that VAX on a single chip. So For me, my inspiration was to work on RISC project

SDF: So at that time what was the implementation technology?

DP: It was NMOS. But Intel was building microprocessors. By the time microprocessors were like toy computers, you could not do real computing on them. So they were hobbies computers, the very first introduction of PCs was just happening at that time

SDF: How about the instruction set design of RISI-I?

DP: The instruction set design at that time was to trying to use quantities principle to design instruction set that would result in good performance and good match to silicon

SDF: From physical technology perspective, Moore's Law still applies in the coming decade most likely, once semiconductor technology hits its dead-end, what is the most promising technology for digital computation¡¯s physical implementation?

DP: I think it is risky to think that silicon technology comes to the dead-end. I have a colleague, Prof. Chenming Hu, he works on circuit design, he also is the CTO at TSMC, and so he is a real world expert. He thinks , things will flow down some, he can imagine that some of the techniques people are using, trying to build Carbon-nano or other nano-related technologies in general, that could be a point to standard transistor technology to make them shrink as well. Speaking as a computer designer, I think, unquestionably, the big opportunity is to figure out how to use lots of independent computers on a chip. Today, if I think about the RISC design you talked about it, about 45,000 transistors, 2,000 of those on a fifty-nanometer chip, including caches and so on. But we do not know how to program it, so we got resources to do that. So I think, what it makes sense, from a computer science perspective, is how to create software and design hardware that can efficiently use thousands of parallel processors efficiently on a chip. If you can do that, even if Moore¡¯s law slows down, we could still figure out more performance, or even some of other technology seems like to me, take advantage of parallel computing. So my idea would be, even Moor¡¯s law is ending, there is important problem to solve. Whatever replaces CMOS, I think it will be able to take advantage of lots of simple computers rather than a big computer

SDF: What's the rule of thumb in designing a good ISA? How does ISA interact with compiler technology and hardware architecture? Say, let¡¯s take RSIC-I as an example

DP: A long time ago, people programmed primarily in assembly language, and someone is still doing that. But as programs got big, that just became infeasible to write a whole program in assembly language. As a result, the emphasis on instruction set design was sufficient match what compiler could produce rather than what human could write. So that was a very important interface in the RISC era. It did make sense to design instruction set that your compiler can use, or it did make sense to design instruction set that can generate code faster than you think. That was a very important and that was the idea behind that.

SDF: According to a presentation made by Prof. A. R. Newton at Berkley, he listed 11 top R & D projects at Berkeley, 3 of them were done by your group, namely RISC, RAID and NOW. Why did you start to work on RAID (Redundant Arrays of Inexpensive Disks) technology after RISC-I project?

DP: What we thought at the time, the processor would get a lot faster, but it was not clear that I/O would improve as fast as that the processing would. So the speed of access, putting something into the storage, if that did not get any faster, we spent a lot of time waiting the storage system even the processor gets a lot faster. So after a while, it would be a good computer if the storage could not get any faster, and that was the original idea. And my colleague, Randy Katz, he just bought a Mac, a desktop computer, a disk came with the Mac. He asked himself the question: this small disk, this is a very interesting piece of technology, what we could do with this k. The combination of the opportunity of small disks, and the need to create the I/O system that can keep with the RISC processing technology, led to the great project .The original theme was how to get more performance by replacing a big disk with lots of small disks

SDF: Why parallelism, scalability and programmability are the top three critical factors in high performance computing? Why storage and communication also become more and more important in high performance computing?

DP: I think what¡¯s called super-computing, was originally focused just on computation. So because that was the hard problem, in beginning, people focused just on that. Now if people want to solve real problems, not just to do research papers on showing how to make program be in parallel and fast. Big problem has big data as well as lots of computations. So once again, like the original RAID argument, if you just did the computation part much fast, and did not anything fast on the storage part, you will be limited by how fast you do the storage. So that¡¯s what happening in high-performance computing, is that you need to have focusing more on storage today more than 10 or 20 years ago, we need to focus more on to try to get high speed storage.

Back to Home Page